Adversarial examples constitute a threat of great seriousness against the stability of neural networks. Adversarial examples exploit the weakness of a model by appending imperceptible but small disturbances that lead to incorrect predictions. This work provides a general setup for generating, detecting, and defending adversarial examples using an aggressive Projected Gradient Descent (PGD) approach. The adopted methodology focuses on the efficacy of adversarial attacks and confirms the efficacy of a perturbation-based detection approach to detect the manipulated input. Adversarial training as a defensive technique is employed, proving competent in recovering robustness to models. The results are a staggering decrease in accuracy from 96.33% clean data to 62.33% adversarial examples, confirming the astounding effect of the attacks. It was noted that 100% detection was seen when a perturbation-based system detected adversarial inputs. Adversarial training lessened such an impact, as accuracy was enhanced to 94.33% clean data and 95.33% for adversarial examples. The study emphasizes the importance of adding detection and mitigation processes to defend neural networks against adversarial settings, which provides innovation opportunities in machine learning resilience applications.